caffe2/operators/reduction_ops.h

/**
 * Copyright (c) 2016-present, Facebook, Inc.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#ifndef CAFFE2_OPERATORS_REDUCTION_OPS_H_
#define CAFFE2_OPERATORS_REDUCTION_OPS_H_

#include "caffe2/core/common_omp.h"
#include "caffe2/core/context.h"
#include "caffe2/core/logging.h"
#include "caffe2/core/operator.h"
#include "caffe2/utils/math.h"

namespace caffe2 {

template <typename T, class Context>
class SumElementsOp : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;

  SumElementsOp(const OperatorDef& operator_def, Workspace* ws)
      : Operator<Context>(operator_def, ws),
        average_(OperatorBase::GetSingleArgument<bool>("average", false)) {}
  SumElementsOp(const OperatorDef& operator_def, Workspace* ws, bool average)
      : Operator<Context>(operator_def, ws), average_(average) {}
  ~SumElementsOp() {}

  bool RunOnDevice() override
// TODO: T21635002 fix float-divide-by-zero undefined behavior
#if defined(__has_feature)
#if __has_feature(__address_sanitizer__)
      __attribute__((__no_sanitize__("float-divide-by-zero")))
#endif
#endif
  {
    auto& X = Input(0);
    auto* sum = Output(0);
    sum->Resize(vector<TIndex>());
    T* data = sum->template mutable_data<T>();
    math::Sum<T, Context>(
      X.size(), X.template data<T>(), data, &context_, &scratch_);
    if (average_) {
      math::Scale<T, Context>(
          1,
          static_cast<T>(1.) / X.size(),
          sum->template data<T>(),
          data,
          &context_);
    }
    return true;
  }

 private:
  bool average_;
  Tensor<Context> scratch_;
};

template <typename T, class Context>
class SumElementsGradientOp : public Operator<Context> {
 public:
  USE_OPERATOR_CONTEXT_FUNCTIONS;

  SumElementsGradientOp(const OperatorDef& operator_def, Workspace* ws)
      : Operator<Context>(operator_def, ws),
        average_(OperatorBase::GetSingleArgument<bool>("average", false)) {}
  SumElementsGradientOp(
      const OperatorDef& operator_def,
      Workspace* ws,
      bool average)
      : Operator<Context>(operator_def, ws), average_(average) {}
  ~SumElementsGradientOp() {}

  bool RunOnDevice() override;

 private:
  bool average_;
};

template <class Context>
class SumSqrElementsOp : public Operator<Context> {
 public:
  USE_SIMPLE_CTOR_DTOR(SumSqrElementsOp)
  USE_OPERATOR_CONTEXT_FUNCTIONS;

  bool RunOnDevice() override {
    return DispatchHelper<TensorTypes<float>>::call(this, Input(0));
  }

  template <typename T>
  bool DoRunWithType() {
    bool average = OperatorBase::GetSingleArgument<bool>("average", false);
    auto& X = Input(0);
    auto* sum = Output(0);
    sum->Resize(vector<TIndex>());
    math::SumSqr<T, Context>(
        X.size(),
        X.template data<T>(),
        sum->template mutable_data<T>(),
        &context_,
        &scratch_);
    if (average) {
      math::Scale<T, Context>(
          1,
          float(1.) / X.size(),
          sum->template data<T>(),
          sum->template mutable_data<T>(),
          &context_);
    }
    return true;
  }

 private:
  Tensor<Context> scratch_;
};

template <typename T, class Context, bool ROWWISE>
class MaxReductionOp : public Operator<Context> {
 public:
  USE_SIMPLE_CTOR_DTOR(MaxReductionOp)
  USE_OPERATOR_CONTEXT_FUNCTIONS;

  bool RunOnDevice() override {
    auto& X = Input(0);
    CAFFE_ENFORCE_EQ(X.ndim(), 3);

    const int batch_size = X.dim32(0);
    const int M = X.dim32(1);
    const int N = X.dim32(2);

    auto* Y = Output(0);
    ROWWISE ? Y->Resize(batch_size, M) : Y->Resize(batch_size, N);

    if (ROWWISE) {
      math::RowwiseMax<T, Context>(
          batch_size * M,
          N,
          X.template data<T>(),
          Y->template mutable_data<T>(),
          &context_);
    } else {
      const int input_size = N * M;
      for (int i = 0; i < batch_size; ++i) {
        math::ColwiseMax<T, Context>(
            M,
            N,
            X.template data<T>() + i * input_size,
            Y->template mutable_data<T>() + i * N,
            &context_);
      }
    }
    return true;
  }
};

template <typename T, class Context, bool ROWWISE>
class MaxReductionGradientOp : public Operator<Context> {
 public:
  USE_SIMPLE_CTOR_DTOR(MaxReductionGradientOp)
  USE_OPERATOR_CONTEXT_FUNCTIONS;

  bool RunOnDevice() override;
};

} // namespace caffe2

#endif